PARAMETER ESTIMATION FOR THE LOG-LOGISTIC DISTRIBUTION BASED ON ORDER STATISTICS Authors: Mohammad Ahsanullah { Department of Management Sciences, Rider University, New Jersey, USA ([email protected]) Ayman Alzaatreh { Department of Mathematics and Statistics, American University of Sharjah, Sharjah, UAE ([email protected]) Abstract: • In this paper, we discuss the moments and product moments of the order statistics in a sample of size n drawn from the log-logistic distribution. We provide more compact forms for the mean, variance and covariance of order statistics. Parameter estimation for the log-logistic distribution based on order statistics is studied. In particular, best linear unbiased estimators (BLUEs) for the location and scale parameters for the log-logistic distribution with known shape parameter are studied. Hill estimator is proposed for estimating the shape parameter. Key-Words: • Log-logistic distribution; moments; order statistics; best linear unbiased estimators; recurrence relations; Hill estimator. AMS Subject Classification: • 62E10, 62F10. 2 Mohammad Ahsanullah and Ayman Alzaatreh Parameter estimation for the log-logistic distribution 3 1 INTRODUCTION The probability density function (pdf) of the log-logistic distribution with unit scale parameter is given by α xα−1 (1.1) f(x) = ; x ≥ 0; (1 + xα)2 where α is a positive real number. A random variable X that follows the density function in (1.1) is denoted as X ∼ log-logistic(α). The cumulative distribution (cdf) and quantile functions of the log-logistic distribution, respectively, are xα (1.2) F (x) = ; x ≥ 0: 1 + xα and x 1/α (1.3) F −1(x) = ; 0 < x < 1: 1 − x The kth moments of the log-logistic distribution in (1.1) can be easily computed as k k (1.4) µ0 = B 1 − ; 1 + ; k α α where B(:; :) is the beta function. Note that the kth moment exists iff α > k: A more compact form of (1.4) can be derived using the fact that Γ(z) Γ(1 − z) = π csc (π z) (Abramowitz and Stegun, 1964) as follows kπ kπ (1.5) µ0 = Γ(1 − k/α) Γ(1 + k/α) = csc ; α > k: k α α Therefore, E(X) = (π/α) csc (π/α) and V ar(X) = (π/α)f2 csc (2π/α)−(π/α) csc2 (π/α)g: The log-logistic distribution is a well-known distribution and it is used in different fields of study such as survival analysis, hydrology and economy. For some applications of the log-logistic distribution we refer the reader to Shoukri et al.[23], Bennett [10], Collet [11] and Ashkar and Mahdi [7]. It is also known that the log-logistic distribution provides good approximation to the normal and 4 Mohammad Ahsanullah and Ayman Alzaatreh the log-normal distributions. The log-logistic distribution has been studied by many researchers such as Shah and Dave [22], Tadikamalla and Johnson [24], Ragab and Green [21], Voorn [25] and Ali and Khan [4]. Ragab and Green [21] studied some properties of the order statistics from the log-logistic distribution. Ali and Khan [4] obtained several recurrence relations for the moments of order statistics. Voorn [25] characterized the log-logistic distribution based on extreme related stability with random sample size. In this paper, we discuss the moments of order statistics for the log-logistic distribution. We review some known results and provide a more compact expression for calculating the covariance between two order statistics. Also, we discuss the parameter estimation of the log-logistic distribution based on order statistics. 2 SOME RESULTS FOR THE MOMENTS OF OR- DER STATISTICS Let X1;X2; :::; Xn be n independent copies of a random variable X that follows log-logistic(α): Let X1;n ≤ X2;n ≤ ::: ≤ Xn;n be the corresponding order statistics. Then from (1.1) and (1.2), the pdf of the rth order statistics is given by α xαr−1 (2.1) f (x) = C ; x ≥ 0; r:n r:n (1 + xα)n+1 n! where Cr:n = (r−1)!(n−r)! : th The k moments of Xr:n can be easily derived from (2.1) as k k (2.2) α(k) = C B n − r + 1 − ; r + ; α > k; r:n r:n α α Similarly as in (1.5), one can show that kπ n (−1)rπ csc Y k (2.3) α(k) = α i − r − ; α > k: r:n (r − 1)!(n − r)! α i=1 (k) k k Note that when r = n = 1; α1:1 = B 1 − α ; 1 + α which agrees with (1.4). From (2.2), the first and second moments of Xr:n are, respectively, given by π n (−1)rπ csc Y 1 (2.4) α(1) = α i − r − ; α > 1; r:n (r − 1)!(n − r)! α i=1 Parameter estimation for the log-logistic distribution 5 and 2π n (−1)rπ csc Y 2 (2.5) α(2) = α i − r − ; α > 2: r:n (r − 1)!(n − r)! α i=1 It is interesting to note that (2.3) can be used easily to derive several re- currence relations for the moments of order statistics. Some of these recurrence relations already exist in the literature. Below, we provide some of these recur- rence relations. I. From (2.3), we can write kπ n−1 −1 (−1)r−1π csc Y k α(k) = α i − (r − 1) − r:n r − 1 (r − 2)!(n − r)! α i=0 kπ n−1 r − 1 + k/α (−1)r−1π csc Y k = α i − (r − 1) − r − 1 (r − 2)!(n − r)! α i=1 k (2.6) = 1 + α(k) ; 2 ≤ r ≤ n: α(r − 1) r−1:n−1 Note that the recurrence relation in (2.6) was first appeared in Ragab and Green (1984). II. If r = 1 in (2.3), then kπ n−1 (k) −π csc k Y k α = α n − 1 − i − 1 − 1:n (n − 1)(n − 2)! α α i=1 k (2.7) = 1 − α(k) ; n ≥ 2: α(n − 1) 1:n−1 The recurrence relation in (2.7) first appeared in Ali and Khan (1987). III. For m 2 N, (2.3) implies r k n (−1) π csc − m π Y k α(k−mα) = α i − r − + m r:n (r − 1)!(n − r)! α i=1 kπ n (r − m − 1)!(n − r + m)! (−1)r−mπ csc Y k = α i − (r − m) − (r − 1)!(n − r)! (r − m − 1)!(n − r + m)! α i=1 (r − m − 1)!(n − r + m)! = (2.8) α(k) ; m + 1 ≤ r ≤ n: (r − 1)!(n − r)! r−m:n When m = 1, (2.8) reduces to the recurrence relation given by Ali and (k−α) n−r+1 (k) Khan (1987) as αr:n = r−1 αr−1:n; 2 ≤ r ≤ n: 6 Mohammad Ahsanullah and Ayman Alzaatreh IV. Another form of (2.8) can be derived as follows kπ n (−1)r+mπ csc Y k α(k−mα) = α i + m − r − r:n (r − 1)!(n − r)! α i=1 kπ n+m (−1)r+mπ csc Y k = α i − r − (r − 1)!(n − r)! α i=m+1 (−1)m Qn+m i − r − k (2.9) = i=n+1 α α(k); m + 1 ≤ r ≤ n: Qm k r:n i=1 i − r − α V. From (2.8) and (2.9), we get (2.10) (−1)m (r − m − 1)!(n − r + m)! Qm i − r − k α(k) = i=1 α α(k) ; m+1 ≤ r ≤ n: r:n Qn+m k r−m:n (r − 1)!(n − r)! i=n+1 i − r − α 3 COVARIANCE BETWEEN ORDER STATISTICS To calculate the covariance between Xr:n and Xs:n, consider the joint pdf of Xr:n and Xs:n, 1 ≤ r < s ≤ n as follows xαr−1yα−1(yα − xα)s−r−1 (3.1) f (x; y) = α2 C ; 0 ≤ x ≤ y < 1; r;s:n r;s:n (1 + xα)s(1 + yα)n−r+1 n! where Cr;s:n = (r−1)!(s−r−1)!(n−s)! : Therefore the product moments, αr;s:n = E(Xr:nYs:n); can be written as Z 1 Z y αr α α α s−r−1 2 x y (y − x ) (3.2) αr;s:n = α Cr;s:n α s α n−r+1 dxdy: 0 0 (1 + x ) (1 + y ) On using the substitution u = xα and v = yα, (3.2) reduces to 1 1 v r+ 1 −1 s−r−1 ! Z v α Z u α (v − u) (3.3) αr;s:n = Cr;s:n n−r+1 s du dv: 0 (1 + v) 0 (1 + u) | {z } I u By using the substitution t = v , it is not difficult to show that I can be simplified to s+ 1 −1 1 1 1 (3.4) I = v α B r + ; s − r F s; r + ; s + ; −v ; α 2 1 α α where pFq is the generalized hypergeometric function defined as 1 k X (a1)k ::: (ap)k x pFq(a1; : : : ; ap; b1; : : : ; bq; x) = : (b1)k ::: (bq)k n! k=0 Parameter estimation for the log-logistic distribution 7 −a x Using the Pfaff transformation, 2F1(a; b; c; x) = (1 − x) 2F1(a; c − b; c; x−1 ); we have 1 1 −r− 1 1 1 1 v (3.5) F s; r + ; s + ; −v = (1 + v) α F ; r + ; s + ; : 2 1 α α 2 1 α α α 1 + v v Now, using (3.4), (3.5) and the substitution w = 1+v , (3.3) reduces to (3.6) Z 1 1 s+ 2 −1 n−s− 1 1 1 1 αr;s:n = Cr;s:n B r + ; s − r w α (1−w) α 2F1 ; r + ; s + ; w : α 0 α α α On using the identity [Gradshteyn and Ryzhik, [14], p.813] Z 1 ρ−1 σ−1 x (1 − x) 2F1(α; β; γ; x)dx = B(ρ, σ)3F2(α; β; ρ; γ; ρ + σ; 1); 0 the product moments of the log-logistic distribution can be written as (3.7) 1 2 1 1 1 2 1 1 α = C B r + ; s − r B s + ; n − s − + 1 F ; r + ; s + ; s + ; n + + 1; 1 : r;s:n r;s:n α α α 3 2 α α α α α It is clear from (3.7) that αr;s:n exists for all α > 1: It is noteworthy to mention that one can use some existing recurrence relations in the literature to compute αr;s:n in a more efficient way.